Re: [PATCH RFC] vfio: Documentation for the migration region

[Cornelia Huck <cohuck@redhat.com>]

On Fri, Nov 26 2021, Jason Gunthorpe <jgg@nvidia.com> wrote:

> On Fri, Nov 26, 2021 at 01:56:26PM +0100, Cornelia Huck wrote:
>> On Thu, Nov 25 2021, Jason Gunthorpe <jgg@nvidia.com> wrote:
>> 
>> > On Thu, Nov 25, 2021 at 01:27:12PM +0100, Cornelia Huck wrote:
>> >> On Wed, Nov 24 2021, Jason Gunthorpe <jgg@nvidia.com> wrote:
>> >> 
>> >> > On Wed, Nov 24, 2021 at 05:55:49PM +0100, Cornelia Huck wrote:
>> >> 
>> >> >> What I meant to say: If we give userspace the flexibility to operate
>> >> >> this, we also must give different device types some flexibility. While
>> >> >> subchannels will follow the general flow, they'll probably condense/omit
>> >> >> some steps, as I/O is quite different to PCI there.
>> >> >
>> >> > I would say no - migration is general, no device type should get to
>> >> > violate this spec.  Did you have something specific in mind? There is
>> >> > very little PCI specific here already
>> >> 
>> >> I'm not really thinking about violating the spec, but more omitting
>> >> things that do not really apply to the hardware. For example, it is
>> >> really easy to shut up a subchannel, we don't really need to wait until
>> >> nothing happens anymore, and it doesn't even have MMIO. 
>> >
>> > I've never really looked closely at the s390 mdev drivers..
>> >
>> > What does something like AP even do anyhow? The ioctl handler doesn't
>> > do anything, there is no mmap hook, how does the VFIO userspace
>> > interact with this thing?
>> 
>> For AP, the magic is in the hardware/firmware; the vfio parts are needed
>> to configure what is exposed to a given guest, not for operation. Once
>> it is up, the hardware will handle any instructions directly, the
>> hypervisor will not see them. (Unfortunately, none of the details have
>> public documentation.) I have no idea how this would play with migration.
>
> That is kind of what I thought..
>
> VFIO is all about exposing a device to userspace control, sounds like
> the S390 drivers skipped that step.

Note that what I wrote above is about AP; CCW does indeed trigger
operations like start subchannel from userspace and relays interrupts
back to userspace. AP is just very dissimilar to basically anything
else.

>
> KVM is all about taking what userspace can already control and giving
> it to a guest, in an accelerated way.
>
> Making a bypass where a KVM guest has more capability than the user
> process because VFIO and KVM have been directly coupled completely
> upends the whole logical model.
>
> As we talked with Intel's wbinvd stuff you should have a mental model
> where the VFIO userspace process can do anything the KVM guest can do
> via ioctls on the mdev. KVM is just an accelerated way to do that same
> stuff. Maybe S390 doesn't implement those ioctls, but they are
> logically part of the model.

FWIW, AP had been a pain to model in a way that we could hand the
devices to the guest; if we are supposed to use vfio for this purpose,
the current design is probably the best we can get, at least nobody has
been able to come up with a better way to interact with the interfaces
that we have.

CCW needs a kernel part for translations, as it doesn't have an iommu,
and the I/O instructions are of course privileged (but so are the
instructions for s390 PCI); I think it is quite close to other devices
in other respects, only that it has a more transaction-based model.

> So, for the migration doc, imagine some non-accelerated KVM that was
> intercepting the guest operations and calling the logical ioctls on
> the mdev instead. When we talk about MMIO/PIO/etc it also includes
> mdev operation ioctls too, and by extension any ioctl accelerated
> inside KVM.

I think only AP is the really odd one out here; CCW will likely differ
in some details... I just wanted to make sure that this will not run
counter to the documentation.